Clustering brain morphometry in Autism Spectrum Disorder (ASD)
Summary
Background: Heterogeneity in Autism Spectrum Disorder (ASD) is a major obstacle for research and clinical practice. This study aimed to identify distinct neuroanatomical subtypes by applying advanced unsupervised clustering to a large, multi-site structural magnetic resonance imaging (sMRI) dataset.
Methods: We analyzed sMRI data from 1,449 male individuals with ASD and 1,285 typically developing (TD) controls from the ENIGMA consortium. We applied robust harmonization (ComBat) and normative modeling to generate individual-level neuroanatomical deviation scores (Z-scores). These features were then subtyped by multiple clustering algorithms (HDBSCAN, HYDRA, NotTooDeep) and evaluated for stability, internal validity, and clinical relevance. Linear Mixed-Effects models were used to assess group-level differences and age-related trajectories.
Results: Despite a rigorous methodological approach, our main finding is a failure to identify any stable, reproducible, or clinically meaningful neuroanatomical subgroups. The most robust clustering solution partitioned the data based on a global brain size effect, and these data-driven groups showed no association with clinical measures. However, group-level analyses did confirm subtle but significant neuroanatomical differences between the ASD and TD cohorts and revealed a complex pattern of age-related changes, with regional brain differences both attenuating and diverging over development.
Conclusion: These findings are strong evidence that the neuroanatomical heterogeneity in ASD is continuous rather than categorical. The search for discrete subtypes using unimodal sMRI data may be a limited approach. Future research should pivot to dimensional models with multimodal data that map continuous brain variability onto the clinical spectrum of autism.